Ixamed’s submission description for WMT20 Biomedical shared task: benefits and limitations of using terminologies for domain adaptation

Proceedings of the 5th Conference on Machine Translation (WMT), pages 875–880 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 875 Ixamed’s submission description for WMT20 Biomedical shared task: benefits and limitations of using terminologies for domain adaptation Xabier Soto, Olatz Perez-de-Vi˜naspre, Gorka Labaka, Maite Oronoz HiTZ Basque Center for Language Technologies - Ixa, University of the Basque Country UPV/EHU {xabier.soto, olatz.perezdevinaspre, gorka.labaka, maite.oronoz }@ehu.eus Abstract In this paper we describe the systems devel- oped at Ixa for our participation in WMT20 Biomedical shared task in three language pairs, en-eu, en-es and es-en. When defining our ap- proach, we have put the focus on making an efficient use of corpora recently compiled for training Machine Translation (MT) systems to translate Covid-19 related text, as well as reusing previously compiled corpora and de- veloped systems for biomedical or clinical do- main. Regarding the techniques used, we base on the findings from our previous works for translating clinical texts into Basque, making use of clinical terminology for adapting the MT systems to the clinical domain. However, after manually inspecting some of the outputs generated by our systems, for most of the sub- missions we end up using the system trained only with the basic corpus, since the systems including the clinical terminologies generated outputs shorter in length than the correspond- ing references. Thus, we present simple base- lines for translating abstracts between English and Spanish (en/es); while for translating ab- stracts and terms from English into Basque (en-eu), we concatenate the best en-es system for each kind of text with our es-eu system. We present automatic evaluation results in terms of BLEU scores, and analyse the effect of in- cluding clinical terminology on the average sentence length of the generated outputs. Fol- lowing the recent recommendations for a re- sponsible use of GPUs for NLP research, we include an estimation of the generated CO2 emissions, based on the power consumed for training the MT systems. 1 Introduction The WMT20 Biomedical shared task calls for de- veloping systems for translating biomedical ab- stracts and terminologies between several lan- guages. In our case, we participate in the task of translating biomedical terms and abstracts from English into Basque (en-eu), as well as translating biomedical abstracts between English and Spanish (en-es and es-en). For translating the test data from English into Basque, we concatenate our best en-es system with our es-eu system, both for translating abstracts and terminologies. 2 Related work For translating biomedical texts from English into Catalan, Costa-juss´a et al. (2018) use a pivoting or cascade approach, translating the texts first from English into Spanish (en-es), and then from Span- ish into Catalan (es-ca). This technique is useful when there are more bilingual in-domain sentences for each of the language pairs (en/es and es/ca) than for the desired source and target languages (en/ca). Since there are low resources for en/eu biomedical domain, but we have access to many resources for en/es and es/eu in the biomedical or clinical do- main, we follow the …

Content truncated to 500 words